Paper deep dive

AURORA: A Natural Language-Driven Agentic Framework for Understanding, Reasoning, and Orchestrating Reliable Air-Ground Co-Simulation

Keshu Wu, Hao Zhang, Rui Gan, Xiangbo Gao, Xiaopeng Li, Zhengzhong Tu, Yang Zhou

Year: 2026Venue: arXiv preprintArea: cs.ROType: PreprintEmbeddings: 63

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%

Last extracted: 9/18/2026, 4:20:43 AM

Summary

The paper introduces AURORA, a natural language-driven agentic framework for generating and verifying air-ground co-simulation scenarios. It utilizes an Air-Ground Scenario Graph (AGSG) to link agents, missions, and events, enabling joint road-airspace grounding, temporal planning, and runtime verification. AURORA-Bench is presented as a benchmark to evaluate scenario realization fidelity. The system treats scenario generation as a compilation process with verification, allowing for localized repair of violations without full regeneration.

Entities (7)

AURORA · framework · 98%Air-Ground Scenario Graph · data-structure · 97%AURORA-Bench · benchmark · 96%AirSim · simulator · 95%CARLA · simulator · 95%UAV · agent-type · 92%LLM · technology · 90%

Relation Signals (6)

AURORA → uses → Air-Ground Scenario Graph

confidence 98% · Central to AURORA is the Air-Ground Scenario Graph (AGSG), a typed intermediate representation...

AURORA → evaluateswith → AURORA-Bench

confidence 96% · We further introduce AURORA-Bench to evaluate not only whether generated scenarios execute...

AURORA → integrates → CARLA

confidence 95% · Synchronized CARLA–AirSim execution provides requirement-level measurements...

AURORA → integrates → AirSim

confidence 95% · Synchronized CARLA–AirSim execution provides requirement-level measurements...

AURORA → manages → UAV

confidence 92% · AURORA supports three core UAV mission modes—hover, track, and waypoint navigation...

AURORA → employs → LLM

confidence 90% · The language model is therefore used for semantic interpretation and guarded edit proposals...

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Air-ground transportation research increasingly relies on co-simulation, yet constructing scenarios remains labor-intensive and difficult to validate. More importantly, a generated scenario may execute successfully while failing to realize the spatial, temporal, communication, or behavioral relationships requested by the user. This paper presents AURORA, a natural-language-driven agentic framework that treats air-ground scenario generation as a process of compilation with verification. Central to AURORA is the Air-Ground Scenario Graph (AGSG), a typed intermediate representation that explicitly connects agents, aerial missions, events, communication links, success conditions, and their cross-domain dependencies. This shared representation enables simulator-grounded parsing, joint road-airspace grounding, temporal planning, pre-execution feasibility checking, trace-based runtime verification, failure localization, and bounded repair within a unified workflow. We further introduce AURORA-Bench to evaluate not only whether generated scenarios execute, but whether they faithfully realize the requested interactions. Experiments across multiple language models show that structured execution substantially improves reliability, while runtime verification exposes silent failures that completion-based evaluation overlooks. Localized repair further resolves many violations without regenerating the entire scenario. The results show that reliable scenario generation requires verifying realized behavior, not merely executable code, and demonstrate the value of explicit intermediate representations for verifiable and repairable language-driven co-simulation.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

59,388 characters extracted from source content.

Expand or collapse full text

AURORA: A Natural Language–Driven Agentic Framework for Understanding, Reasoning, and Orchestrating Reliable Air–Ground Co-Simulation Keshu Wu 1 , Hao Zhang 1 , Rui Gan 2 , Xiangbo Gao 3 , Xiaopeng Li 2 , Zhengzhong Tu 3 , and Yang Zhou 1,† Abstract— Air–ground transportation research increasingly relies on co-simulation, yet constructing scenarios remains labor-intensive and difficult to validate. More importantly, a generated scenario may execute successfully while failing to realize the spatial, temporal, communication, or behavioral rela- tionships requested by the user. This paper presents AURORA, a natural-language-driven agentic framework that treats air– ground scenario generation as a process of compilation with verification. Central to AURORA is the Air–Ground Scenario Graph (AGSG), a typed intermediate representation that ex- plicitly connects agents, aerial missions, events, communication links, success conditions, and their cross-domain dependencies. This shared representation enables simulator-grounded pars- ing, joint road–airspace grounding, temporal planning, pre- execution feasibility checking, trace-based runtime verifica- tion, failure localization, and bounded repair within a unified workflow. We further introduce AURORA-Bench to evaluate not only whether generated scenarios execute, but whether they faithfully realize the requested interactions. Experiments across multiple language models show that structured execu- tion substantially improves reliability, while runtime verifica- tion exposes silent failures that completion-based evaluation overlooks. Localized repair further resolves many violations without regenerating the entire scenario. The results show that reliable scenario generation requires verifying realized behavior, not merely executable code, and demonstrate the value of explicit intermediate representations for verifiable and repairable language-driven co-simulation. I. INTRODUCTION Autonomous driving, intelligent transportation systems, and the emerging low-altitude economy are creating mobility systems in which unmanned aerial vehicles (UAVs) [1], roadside infrastructure [2], connected vehicles [3], [4], and other traffic participants [5] must perceive, communicate, and coordinate within a shared environment. Air–ground simula- tion [6], [7], [8] provides a practical testbed for studying such interactions before deployment, particularly for applications such as cooperative perception, emergency response, and connected transportation. Yet constructing such experiments 1 Keshu Wu, Hao Zhang, and Yang Zhou are with the Zachry De- partment of Civil and Environmental Engineering, Texas A&M Uni- versity, College Station, TX, USA.keshuw, zhanghao230, yangzhou295@tamu.edu 2 Rui Gan and Xiaopeng Li are with the Department of Civil and Environmental Engineering, University of Wisconsin–Madison, Madison, WI, USA.rgan6, xli2485@wisc.edu 3 Xiangbo Gao and Zhengzhong Tu are with the Department of Computer Science and Engineering, Texas A&M University, College Station, TX, USA.xiangbog, tzz@tamu.edu † Corresponding author: Yang Zhou, yangzhou295@tamu.edu remains labor-intensive. Researchers must manually coor- dinate map selection, valid actor placement, UAV routing, simulator synchronization, event triggers, communication links, and success criteria, and these steps must be repeated for each new scenario. As a result, the scale and diversity of air–ground experiments are often limited not by simulator capability, but by the effort required to author and validate scenarios. Natural-language and generative-AI interfaces offer a promising way to reduce this burden and improve interaction with connected and automated transportation systems [9], [10], [11]. Recent work has explored language-driven traffic scenario generation and editing [12], [13], [14], [15] and natural-language UAV control in simulation [16], [17], [18]. However, air–ground scenario generation requires more than translating a prompt into executable code. A scenario can run while violating requested spatial, temporal, perceptual, or communication relations—for example, through invalid placement, premature mission transitions, missed messages, or interactions that never occur. Such failures are difficult to detect when generated scripts lack an explicit specification of what should happen. Specification-based simulation anal- ysis and safety evaluation provide foundations for detecting such failures [19], [20], but air–ground scenarios addition- ally require joint feasibility across road geometry, three- dimensional airspace, sensing, communication, and event timing. The challenge is therefore not only to generate exe- cutable scenarios, but to verify that the requested interactions are realized before and during execution. We present AURORA 1 (A Natural Language–Driven Agentic Framework for Understanding, Reasoning, and Orchestrating Reliable Air–Ground Co-Simulation), which treats language-driven scenario generation as compilation with verification. Natural-language requests are converted into typed, provenance-aware specifications and an Air– Ground Scenario Graph (AGSG) grounded in simulator mea- surements. Deterministic modules then resolve spatial and temporal feasibility, while a managed executor synchronizes the ground and aerial simulators and evaluates trace-level re- quirements. When a violation occurs, the AGSG localizes the implicated components and constrains the repair scope. The language model is therefore used for semantic interpretation and guarded edit proposals, while simulator-bound quantities are grounded and validated through deterministic procedures. 1 Project page: https://keshuw95.github.io/AURORA/ 1 arXiv:2609.19527v1 [cs.RO] 17 Sep 2026 Ascend controlled climb to mission altitude Hover position hold over a target area Track follow a moving ground target Navigate waypoint-guided route traversal Land controlled descent to touchdown Fig. 1. UAV mission primitives in AURORA: ascend, hover, track, navigate, and land. AURORA supports three core UAV mission modes— hover, track, and waypoint navigation—each executed be- tween controlled ascent and landing phases (Fig. 1). To- gether, these primitives cover stationary aerial observation, dynamic air–ground following, and route-based missions while sharing a common execution and verification structure. The key distinction is between execution and realization. AURORA does not treat simulator completion as success; it verifies whether the requested relations and outcomes hold in the execution trace. The AGSG links each requirement to its dependencies, verification conditions, and adjustable parameters, enabling unmet requirements to be localized and repaired without regenerating the full scenario. Verification thus guides both scenario generation and refinement rather than serving as a post hoc check. The contributions are three-fold: 1) AURORA for verified air–ground scenario genera- tion. We formulate natural-language-driven air–ground scenario generation as a verified compilation process and develop an end-to-end agentic workflow that inte- grates simulator-grounded understanding, joint spatial and temporal reasoning, synchronized co-simulation, runtime verification, and bounded repair. 2) Air–Ground Scenario Graph for verification and repair. We introduce the AGSG, a typed interme- diate representation that links requested relations to executable verification conditions, execution evidence, and the parameters that can affect them. This shared representation supports pre-execution feasibility check- ing, trace-based monitoring, failure localization, and localized repair. 3) AURORA-Bench for evaluating scenario real- ization. We construct a dedicated benchmark for language-generated air–ground co-simulation, com- prising 200 prompts, 904 annotated requirements, and boundary-feasible and infeasible cases. A six- configuration, five-LLM evaluation separates execution reliability, prompt fidelity, and runtime realization, revealing failures that completion-based evaluation would otherwise overlook. I. RELATED WORK A. Air–ground simulation and cooperative perception Aerial sensing complements ground-based perception by reducing occlusion and extending coverage. Drone-assisted datasets demonstrate the value of vehicle–UAV collaboration across diverse environments [6], [21], [22], [23], extending prior vehicle–vehicle and vehicle–infrastructure cooperative perception and reasoning [24], [25], [26]. CARLA and Air- Sim provide widely used ground and aerial simulation capa- bilities [27], [28], while newer platforms integrate multirotor and vehicle dynamics [7]; OpenCDA supports cooperative driving automation [29]. Despite these capabilities, scenario construction and cross-domain coordination remain largely manual—valid placement, UAV mission design, synchroniza- tion, event coordination—and AURORA targets this author- ing and verification layer rather than introducing another simulator. B. Scenario specification and language-driven generation Scenic [30] and OpenSCENARIO [31] provide struc- tured traffic-scenario representations, but require special- ized languages or predefined schemas. Learning-based methods further automate traffic generation and plan- ning, including trajectory-generation approaches [32], [33], [34]. Recent language-driven methods, including LLMSce- nario [14], TTSG [12], LCTGen [35], ChatScene [13], and Talk2Traffic [15], generate or interactively edit traffic sce- narios from high-level descriptions, while ScenarioNet [36] provides a data-driven framework for large-scale scenario modeling and generation. These approaches reduce manual specification but primarily target ground traffic and scene or trajectory generation. Air–ground scenarios additionally cou- ple road geometry, three-dimensional flight, sensing, com- munication, and event timing, requiring a representation that captures cross-domain dependencies and joint realizability. C. Language-model execution and scenario verification Language models have also been explored for executable code, embodied control, and transportation reasoning [37], [38], [39], [10], [34], and have recently been applied to natural-language and agentic UAV control in simulated en- vironments [16], [17], [18]. Closed-loop UAV systems have further used execution feedback to evaluate and refine LLM- generated operation code [40]. However, successful execu- tion does not guarantee that a generated scenario satisfies the request, and simulator scripts typically lack an explicit dependency structure for detecting and localizing seman- tic failures. VerifAI [19] demonstrates specification-based simulation analysis, while SafeBench [20] provides sys- tematic benchmarking of autonomous-driving safety. These approaches, however, do not address natural-language com- pilation and verification of coupled air–ground scenarios. 2 Air-Ground Scenario Graph (AGSG) Ground agents Aerial agents Infrastructure Events Evaluation oracles Prompt Understanding & Retrieval Natural-Language Prompt Simulator Knowledge Base Multi-Channel Retrieval Maps Assets APIs Missions Constraints History Semantic Parsing Entities Relations Goals Constraints Comms Time Prompt: ” A drone tracks a truck at an intersection and warns the ego vehicle over degraded V2X." RAG LLM Agent Scenario Reasoning & Verification Joint Road-Airspace Grounding Bind agents and assets to AGSG; revolve frames and semantics Deterministic Execution Planning Behavior Compilation Temporal Scheduling Trigger Synthesis Three-Level Verification L1 Schema & Symbolic Validity Static feasibility and grounding geometry Mission & Semantic Consistency L2 L3 Synchronized Co-Simulation CARLA AirSim Physics (vehicles) Sensors (LiDAR, cameras, IMU) Traffic & Rules Rendering Physics (drones) Sensors (IMU, GPS, cameras) Wind & Weather Rendering Simulation Coordination Aerial Feedback Closed-Loop Minimal Repair Guarded LLM Edit Revise specification/plan under formal guards Localized Optimization Optimize parameters/controls Numeric Adjustment Bounded parameter updates Re-execute ← Re-verify ← Apply Repair Monitoring and Repair Safety & Mission Compliance Runtime Monitoring Compliance Violation Detected if unresolved if unresolved 4 2 5 3 1 Fig. 2. Overview of AURORA. Retrieval-grounded parsing constructs the AGSG for joint road–airspace grounding, temporal planning, and feasibility checks. Synchronized CARLA–AirSim execution provides requirement-level measurements for graph-based failure localization and bounded repair. AURORA bridges these directions through an Air–Ground Scenario Graph that supports feasibility checking, runtime verification, failure tracing, and localized repair. I. METHODS A. Overview and Problem Formulation AURORA converts a natural-language request into an executable and verified air–ground co-simulation. The frame- work receives a prompt p, a simulation environment E specifying the simulator backends and map, and a knowledge base K constructed offline from direct measurements of E . The objective is to generate a scenario whose execution realizes the spatial, temporal, behavioral, and communication requirements expressed in the prompt. Executability and requirement satisfaction are treated separately: a scenario may complete without runtime errors while failing to realize the requested interaction. AURORA represents the request through a typed scenario specification: S = (W,A,u,M,V E ,O),(1) comprising world settings W , ground agents and behav- iors A, UAV mission u, communication links M, trigger– action events V E , and executable success conditions O. The specification separates discrete scenario structure from the quantitative parameters that instantiate it, such as positions, altitudes, distances, timing thresholds, and communication settings. Each parameter also retains provenance indicating whether it is explicit, inferred, or default. This distinction allows later repair to preserve user-specified values more strongly while giving greater flexibility to values introduced by the parser or schema. The generation and verification process is formulated as (S ∗ ,H ∗ ,R ∗ ) = AURORA(p;E,K),(2) whereS ∗ is the returned specification,H ∗ its execution trace, and R ∗ the verification and repair report. A successful output must pass pre-execution feasibility checks and satisfy all encoded success conditions during execution. Rather than returning only a completed simulation, AURORA retains the evidence needed to interpret the outcome, including violated conditions, their satisfaction margins, and any repairs ap- plied. AURORA follows three phases (Fig. 2). During under- standing, retrieval-grounded parsing converts the prompt into a typed specification and its Air–Ground Scenario Graph (AGSG). During reasoning, deterministic modules determine where and when the requested relations can be realized 3 and whether the grounded configuration is feasible before execution. During orchestration, a managed executor syn- chronizes the simulators, records the resulting trace, and evaluates trace-level requirements. Violations are mapped through the AGSG to a constrained set of candidate edits, allowing verification feedback to directly guide subsequent repair. The LLM therefore handles semantic interpretation and guarded symbolic proposals, while simulator-bound grounding, validation, execution, and numerical evaluation remain deterministic. B. Prompt Understanding and Scenario Representation The LLM agent interprets the prompt using the scenario schema and retrieved simulator knowledge, producing a typed specification rather than simulator code. The schema constrains admissible agent types, mission modes, event structures, parameters, units, and references. Invalid outputs are returned to the agent with explicit feedback from schema and static checks, allowing the interpretation to be revised through bounded retries. Accepted fields retain their prove- nance, supporting both prompt-fidelity evaluation and later control over which fields may be modified during repair. Retrieval draws on a knowledge base K built offline from direct simulator measurements and past executions. It con- tains verified spawn locations and map attributes, measured asset geometry, UAV mission templates and operating limits, feasibility constraints, simulator interfaces, and execution history. Core registries remain available throughout gener- ation, while retrieval selects prompt-relevant assets, spawn examples from the active map, and prior episodes with similar requirements. This measurement-grounded knowl- edge reduces dependence on potentially inaccurate model knowledge of simulator APIs or geometry. Importantly, the parser and verifier share the same simulator-grounded records. The knowledge base therefore acts as a common source of constraints: quantities presented to the LLM during interpretation are drawn from the same environment knowledge later used to evaluate feasibility. This alignment reduces inconsistencies in which a value appears plausible during generation but violates the actual simulator configuration during execution. The validated specification is converted deterministically into an Air–Ground Scenario Graph (AGSG): G = (V,E G ,τ V ,τ E , Θ,ω),(3) where V contains environment, ground-agent, UAV, commu- nication, event, and success-condition nodes; E G encodes their relations and dependencies; τ V and τ E assign node and edge types; Θ contains quantitative parameters; and ω maps each parameter to its owning node. The resulting graph explicitly represents dependencies among scenario components—for example, placement affects sensing geome- try, sensing may trigger communication, and communication may enable a subsequent vehicle action. Rather than serving only as a scenario description, the AGSG links requested relations to their verification condi- tions and influencing parameters. Typed dependencies across (a) hover environment ego bus pedestrian 1 drone comm link ego brakes start pedestrian warn ego occlusion fraction met ego stops, gap kept ascend first message delivered ACTORS COMM & EVENTS SUCCESS (b) navigate environment ego drone ascend first ACTORS SUCCESS (c) track environment ego truck 1 drone comm link warn ego start truck min distance reached ascend first message delivered ACTORS COMM & EVENTS SUCCESS environment in environment agent placed relative to drone tracks comm endpoint event triggered by success acts on measures Fig. 3. Representative grounded AGSGs: (a) cooperative occlusion warning; (b) survey patrol; (c) tracking under degraded communication. spatial, temporal, perceptual, communication, and behavioral requirements allow the same graph to support feasibility checking, trace-based monitoring, failure localization, and repair. When a condition fails, its dependencies can be traced to the components and parameters that may have caused the violation. The AGSG therefore connects what the user requests, what the simulator executes, and what can be changed during repair. Fig. 3 shows representative grounded AGSGs across the supported mission modes. C. Joint Grounding, Planning, and Verification The agentic workflow delegates spatial grounding, tem- poral planning, and feasibility checking to deterministic modules. These modules determine where and when the scenario can unfold and return constraint violations to guide refinement. Together, they translate the relational specifica- tion into a configuration that is grounded in the actual road network, airspace, and execution constraints. a) Joint road–airspace grounding: A feasible location must simultaneously support the requested traffic event, UAV mission, sensing geometry, and communication geometry. Candidate road anchors are therefore ranked jointly as s joint (r i ) = αs road (r i )+βs air (r i )+γs sense (r i )+ηs comm (r i ), (4) where the four terms respectively evaluate road suitability, aerial clearance, air-to-ground visibility, and communication geometry. Equal weights are used in the present implementa- tion. The road term characterizes whether the local roadway can support the requested ground interaction, while the remaining terms assess whether the same location provides suitable airspace and air–ground relationships. Once an anchor is selected, ground agents are placed on valid road or pedestrian elements using relative-placement rules. Requested placements that are not directly realizable are projected onto feasible lanes or pedestrian elements while preserving their intended spatial relations as closely as possible. UAV routes are then checked against an obstacle- height lattice, and infeasible waypoints or route segments are corrected before execution. This division keeps structural geometric correction within grounding rather than deferring such defects to the later runtime repair loop. b) Temporal planning: Each UAV mission is executed as a predicate-guarded state machine consisting of ascent, one of three supported mission modes (i.e., hover, track, 4 or waypoint navigation), and landing (Fig. 1). Tasks are scheduled to respect physical feasibility and execution de- pendencies; sequential actions are completed in the required order before subsequent actions begin. For instance, naviga- tion begins only after the UAV reaches its target altitude. Using realized predicates rather than fixed delays prevents downstream actions from beginning before their physical prerequisites are satisfied. Inter-event timing is modeled using a Simple Temporal Network [41], where event orderings and deadlines are ex- pressed as interval constraints. For events e i and e j , temporal relations can be represented as l ij ≤ t e j −t e i ≤ u ij , allowing requirements such as message transmission preceding recep- tion or a warning arriving before a subsequent vehicle action to be checked before execution. After grounding, geometry-dependent triggers are adjusted to ensure that the requested conditions remain reachable from the realized initial configuration. This step is important because a trigger defined against an imagined configuration may become immediately true or unreachable once agents are projected onto valid simulator geometry. Temporal planning therefore operates on the grounded scenario rather than independently of it. c) Three-level verification: Verification proceeds from inexpensive structural checks to trajectory-level require- ments. First, schema verification checks agent and task types, ranges, units, references, and temporal consistency. These checks identify malformed or internally inconsistent specifi- cations before simulator execution. Second, static feasibility checks whether placements, routes, relational constraints, and communication requirements can be realized in the grounded environment. These checks use the instantiated geometry and simulator-grounded limits rather than language-level plausi- bility alone. Violations are associated with the relevant fields or relations so that infeasible configurations can be revised before execution. Third, semantic verification evaluates trace- dependent requirements such as visibility, message delivery, event ordering, and mission completion. These properties depend on interactions over time and therefore cannot be established from the initial configuration alone. They are monitored using quantitative temporal-logic semantics [42], [43], [44], producing a signed margin ρ i : ρ i ≥ 0 indicates satisfaction, whereas ρ i < 0 quantifies the degree of vio- lation. Unlike a binary pass/fail result, the signed margin also indicates how far a condition lies from satisfaction and can therefore guide subsequent correction. The same representation thus supports both verification and repair. D. Synchronized Execution, Monitoring, and Repair a) Managed execution and synchronization: A speci- fication that passes pre-execution checks configures a fixed managed executor for initialization, spawning, control, event scheduling, communication, logging, and cleanup. The LLM defines the scenario specification, while a validated execution path translates that specification into simulator operations. This separation avoids generating new simulator code for each scenario. CARLA and AirSim advance under a shared logical clock. Coordinate transformation and a proxy UAV maintain a common ground–air representation, while online clock correction limits temporal drift. Communication is evaluated from realized sender–receiver geometry together with config- ured loss, latency, occlusion, and retransmission constraints. Agent states, message delivery, and event outcomes are recorded in the execution trace and evaluated against the encoded success conditions. b) Runtime monitoring: Pre-execution feasibility does not guarantee that the requested interaction will occur. Mov- ing agents may fail to establish the intended relation, tracking may not converge, or an event may miss its timing or communication requirement. Runtime monitoring therefore evaluates success conditions on the realized trace and returns their signed robustness margins. This separates verified real- ization from nominal completion and quantifies how far each violated condition lies from satisfaction. c) Bounded local repair: Because trace-level verifica- tion requires co-simulation, AURORA uses bounded local repair: S 0 = Ground Parse K (p) , S k+1 = Repair(S k ,VS k )∈N (S k ), k < B, (5) where VS k = m : ρ m (H k ) < 0 identifies conditions violated by trace H k , and B denotes the maximum number of repair iterations. The AGSG traces each violation to implicated parameters and fields, defining the admissible edit neighborhood N (S k ). This localization prevents unrelated parts of the scenario from changing. Each numerical repair adjusts at most five parameters, while each symbolic repair revises at most six fields and must pass schema and static verification. For coupled numerical failures, a kinematic replay sur- rogate evaluates candidate edits through a provenance- regularized objective: ∆Θ ∗ = argmin ∆Θ∈Ω max 0, ε ⋆ − min m ˆρ m (Θ + ∆Θ) + λ sem X j w(c j ) |∆θ j | width(D j ) . (6) Here, Ω restricts edits to the localized parameter set and schema domains D j ; ˆρ m is the surrogate-estimated satisfac- tion margin; c j denotes the provenance class of parameter θ j ; and ε ⋆ > 0 specifies the desired robustness headroom. The first term drives the worst monitored condition toward a positive margin, while the second penalizes deviation from the parsed request after normalizing edits by their admissible ranges. The coefficient λ sem balances satisfaction against semantic deviation, and w(c j ) penalizes changes to explicit values more heavily than changes to inferred values or defaults. The surrogate screens candidate edits before another full co-simulation run. Repair proceeds from single-parameter adjustment to localized multivariate search, followed by guarded LLM revision when numerical search cannot remove 5 Capability CategoriesRequirement Types a basic (6.7%) spatial (6.2%) sequential (7.2%) conditional (11.7%) coop. perception (17.9%) comm. failure (9.3%) safety-critical (6.3%) patrol & survey (7.3%) multi-agent (11.4%) compound (12.6%) boundary probe (2.1%) infeasible probe (1.2%) agent presence (20.1%) triggered event (16.8%) drone mission (14.8%) success criterion (14.0%) relative placement (10.2%) numeric bound (8.0%) agent behavior (6.1%) communication link (5.9%) environment (4.1%) 12345678910 requirements per prompt 0% 5% 10% 15% 20% 25% share of prompts mean 4.5 f Annotation Density 0%20%40% message delivered ego stops, min gap held ascent before tracking visible to drone occluded from ego agents close below bound tracked fraction 36% 19% 13% 12% 9% 9% 2% e Success Criteria no trigger 46% 1 event 39% 2 events 10% 3 events 6% 200 prompts d Temporal Structure track 42% hover 34% navigate 12% relay-only 12% 200 prompts c Drone Missions 0%10% van pedestrian bus truck police motorcycle ambulance car firetruck 15% 15% 13% 13% 12% 11% 8% 7% 7% b Agent Types drone missions event triggers event actions success criteria agent types placement relations agent behaviors environment settings 0.89 0.90 0.79 0.88 0.98 0.67 0.79 0.91 g Channel Diversity Fig. 4. AURORA-Bench composition, requirement annotations, and scenario diversity. the violation. The LLM receives localized diagnostic feed- back and may edit only within the admissible scope; each proposal must pass deterministic checks before re-execution. Structural geometric failures requiring broader changes, such as an infeasible route, are returned to the grounding stage rather than forced through the minimal-edit repair loop. The loop terminates on success or budget exhaustion. The final report retains verification status, repair history, and unresolved margins. Because provenance weighting discour- ages but does not prohibit changes to explicit values, the report also distinguishes improvements in realized behavior from changes to the monitored conditions themselves. IV. EXPERIMENTAL DESIGN A. Benchmark and Evaluation Protocol AURORA-Bench comprises 200 prompts spanning 12 ca- pability categories and 904 annotated requirements (Fig. 4). Annotations include reference specifications, feasible real- izations where applicable, variation ranges, and paraphrase groups. For prompts without a flight-task requirement, any feasible UAV mode is accepted. End-to-end evaluation uses a 50-prompt subset: 40 behavioral scenarios across nine categories, five boundary-feasible probes, and five infeasible probes. We evaluate five LLMs—GPT-4o, GPT-5.5, GPT-5.4-mini, Gemini 3.1 Pro, and Gemini 3.8 Flash—using provider- default sampling with bridged CARLA 0.9.15 and Air- Sim 1.8.1. All configurations use the same prompts and simulation seed. Within each campaign, one model handles parsing, baseline code generation, and repair proposals. Re- trieval history is frozen before each campaign to prevent information transfer from evaluation runs. Specifications, execution traces, verification reports, and model transcripts are archived. GPT-4o serves as the reference model for the illustrative example and aggregate analysis. B. Metrics and Baselines We distinguish three properties of generated scenarios: executability, prompt fidelity, and runtime realization. Ex- ecutability asks whether a scenario completes without setup errors, runtime errors, or timeouts. Prompt fidelity measures whether the generated specification preserves the annotated request. Runtime realization asks whether the monitored con- ditions actually hold during execution. Accordingly, comple- tion rate (C.) measures successful execution, while runtime- failure rate (RF) captures execution failures. Verified-pass rate (P) measures executions in which all monitored condi- tions are satisfied, whereas silent-failure rate (SF) captures completed runs with unmet requirements. Prompt fidelity (Fid.) measures correspondence between the generated spec- ification and annotated requirements. Verification applies to the encoded monitored conditions, while fidelity and repair logs assess whether parsing and refinement preserve the original request, particularly when thresholds are modified. Supporting metrics characterize grounding accuracy, synchronization, event timing, commu- nication, mission performance, and computational cost. Six configurations progressively introduce framework components: direct code generation (B1), simulator docu- mentation (B2), simulator-grounded retrieval (B3), the typed AGSG and managed executor without online verification (B4), static verification (B5), and runtime monitoring with localized repair (AURORA). Successful B1–B3 runs are classified as completed-unverifiable, as they lack executable specifications for the semantic evaluation used here. Comple- tion is compared across all six configurations, while verified outcomes are compared across B4, B5, and AURORA. V. RESULTS A. Illustrative Example Fig. 5 illustrates how AURORA translates a tracking request into coordinated air–ground behavior. The UAV com- pletes ascent before tracking the bus, while the police–bus interaction triggers a warning to the ego vehicle. The AGSG connects the participating agents, mission, communication link, and event dependencies to the grounded trajectories and execution schedule. These complementary views show not only how the requested interaction develops over time, but also whether its prerequisites and dependent events are realized in the intended order. The warning reaches the ego vehicle before the UAV establishes close tracking. This sequence highlights why 6 !($"! !&!&$"!" (!&%'%% !!($"! !& #$&(&" &$% !#"!& &$$* &%"! %'$% !($"! !& " '% #" $"! " )$!" %&$&'% %&$&#" !!& %&!") %!&"$ &$! %% ($ 212 12 12 12 1 ),3+!2).-),%1 1 1 )11).-(!1% 0.-%4%-21 .,,3-)#!2).- $0.-%,)+%12.-% 5!6/.)-2 20)′%0/+!--%$ 20)′%00%!+)7%$ ,%11!'%1%-2 ,%11!'%$%+)4%0%$ 2!*%.&&!1#%-2$.-% 20!#*)-'12!021 2!0'%2+.#*%$+!-$)-'#,$ +!-$%$ "3112!0218/.+)#%12!021 5!0-%'. $%+)4%0%$ " "" " ! Prompt: “A drone tracks a bus from 40 meters away and alerts the ego vehicle when a police car begins chasing the bus through an intersection.” %! # $ " ! #" !"# #! " ! ! a3D ViewbTop-Down View cAir–Ground Scenario Graph dTask Automaton and Executed ScheduleeExecution time series Fig. 5. Tracking scenario linking the language request, AGSG, grounded trajectories, and monitored outcomes. a basic_hover_02 hover b coop_occlusion_01 hover c basic_track_00 track d multi_00 track e safety_tailgate_01 track f patrol_02 navigate g patrol_03 navigate h patrol_01 navigate ego anchor requested waypoints flown route tracked vehicle ego path other agents road network blocked airspace Fig. 6. Representative grounded hover, track, and navigation scenarios with UAV and ground-agent trajectories and obstacle-aware route correction. communication and mission performance require separate verification: successful message delivery does not establish sustained tracking. The event records confirm the warning outcome, while the UAV–bus distance trace reveals tracking convergence and subsequent variation. A single completion label would collapse these outcomes, whereas requirement- level monitoring preserves their distinction. By linking each condition to execution evidence, the AGSG also provides a basis for targeted diagnosis when part of the requested interaction remains unmet. Fig. 6 broadens the example across the three supported mission modes. Hover scenarios maintain aerial coverage around a grounded region, tracking scenarios couple UAV motion to a moving ground target, and navigation scenarios realize waypoint-based routes within the available airspace. Most requested routes are preserved directly, while geomet- rically infeasible waypoints are corrected during grounding before execution. Together with Fig. 5, these examples illustrate how AURORA translates different language-level mission structures into grounded and executable air–ground scenarios. Fig. 7 shows the executed scenarios from the UAV camera, with one strip per request. The navigation example follows the requested survey route, while the tracking examples keep the motorcycle, bus, and tailgating van in view as the interac- tions unfold. These views provide qualitative evidence that the requested spatial and tracking interactions are realized in simulation, complementing the trace-based verification of communication and event outcomes. B. End-to-End Performance The six configurations across five LLMs isolate the ef- fects of simulator knowledge, structured execution, and verification-driven repair (Table I and Fig. 8). The pro- gression also separates improvements in executability from improvements in realized behavior, which do not necessarily arise from the same framework components. a) Execution reliability: Simulator knowledge im- proves code execution but does not eliminate runtime fail- ures. Documentation raises completion across all five LLMs, 7 t = 6 st = 18 st = 30 st = 43 st = 56 st = 67 s (a) “In clear weather at midday, a drone flies a rectangular survey route at 30 meters altitude over the streets around the ego car.” t = 3 st = 11 st = 18 st = 26 st = 34 st = 40 s (b) “A drone ascends to 40 meters and then follows a motorcycle as it drives through the streets.” t = 3 st = 10 st = 17 st = 25 st = 33 st = 39 s (c) “A bus drives ahead of the ego car and an ambulance follows behind it while two background cars circulate; a drone tracks the bus from 35 meters.” t = 6 st = 14 st = 20 st = 25 st = 45 st = 68 s (d) “A van tailgates the ego car closely, closing from 45 meters; a tracking drone alerts the ego when the van comes within 20 meters.” Fig. 7. Prompt-to-execution examples across four scenarios, showing six onboard UAV views sampled over each execution and the corresponding natural- language request. TABLE I END-TO-END PERFORMANCE ACROSS SIX CONFIGURATIONS AND FIVE LANGUAGE MODELS. BEST VALUE IN EACH COLUMN IS UNDERLINED. GPT-4oGPT-5.5GPT-5.4-miniGemini 3.1 ProGemini 3.8 Flash MethodC.↑ RF↓P↑ SF↓ Fid.↑ C.↑ RF↓P↑ SF↓ Fid.↑ C.↑ RF↓P↑ SF↓ Fid.↑ C.↑ RF↓P↑ SF↓ Fid.↑ C.↑ RF↓P↑ SF↓ Fid.↑ B1: Direct Code Generation 0.42 0.58— 0.82 0.18— 0.70 0.30— 0.80 0.20— 0.72 0.28— B2: + Documentation0.68 0.32— 0.92 0.08— 0.86 0.14— 0.94 0.06— 0.92 0.08— B3: + Grounded Retrieval0.66 0.34— 0.92 0.08— 0.78 0.22— 0.98 0.02— 0.98 0.02— B4: + AGSG & Executor1.00 00.58 0.42 0.966 1.0000.42 0.58 1.0001.0000.50 0.50 1.0001.0000.66 0.34 0.993 1.0000.62 0.38 0.984 B5: + Static Verification0.9800.56 0.42 0.9801.0000.48 0.52 1.0001.0000.60 0.40 0.995 0.9800.70 0.28 0.9951.0000.54 0.46 0.971 AURORA0.9800.760.220.9801.0000.640.361.0001.0000.700.300.995 0.9800.840.140.9951.0000.740.260.971 406080100 GPT-5.5 GPT-5.4 Mini Gemini 3.8 Flash GPT-4o Gemini 3.1 Pro 98 66 98 78 92 a Completed (%) B1B2B3 406080 84 76 74 70 64 b Verified (%) B4B5AURORA Fig. 8. Configuration comparison across models: (a) completion for code- generation baselines; (b) verified passes for AGSG-based configurations. while retrieval provides mixed additional gains, indicating that additional context alone does not guarantee reliable execution. The larger change comes from structured execu- tion: introducing the typed AGSG and managed executor (B4) raises completion to 100% across all models and eliminates observed runtime failures. Yet 34–58% of scenar- ios remain silent failures. Where later configurations have slightly lower completion despite zero runtime failures, the difference reflects pre-execution rejection rather than execu- tion breakdown. Reliable execution therefore does not imply successful realization, making completion alone insufficient for evaluating generated scenarios. b) Verification and repair: Runtime feedback provides the main gain in scenario realization. Static verification produces mixed changes in verified-pass rate, because iden- tifying a feasible configuration before execution does not guarantee that its trace-dependent requirements will hold. Adding runtime monitoring and localized repair improves 8 verified pass 52% silent failure 40% runtime failure 6% false rejection 2% 76% 22% 2% before repair after repair 050%100% basic spatial sequential conditional coop perception comm failure safety critical patrol survey multi agent completedverified (pre-repair) basic spatial sequential conditional cooperative perception communication failure safety- critical patrol / survey multi-agent boundary probes 50% 100% a Outcome Flow b Per-Category Outcome c Verified Pass by Category after repairbefore repairrepair gain Fig. 9. GPT-4o evaluation: (a) repair outcomes; (b) completion and initial verification by category; (c) category-level repair gains. verified-pass rates by 10–20 percentage points over B5 across the five models, reaching 64–84% under AURORA. The gains vary by scenario category (Fig. 9): spatial, sequential, and patrol scenarios improve substantially, while conditional scenarios remain more challenging. This pattern suggests that localized repair is most effective when a violation can be traced to a compact set of geometric, mission, or timing parameters; longer conditional chains remain harder to resolve. Remaining violations are reported explicitly rather than being absorbed into completion. c) Prompt fidelity and realized behavior: Prompt fi- delity and runtime realization capture different properties. GPT-5.5, for example, achieves perfect specification fidelity without achieving the highest verified-pass rate. Preserving the request during parsing is therefore necessary but not sufficient; the grounded geometry, parameters, event tim- ing, and mission dynamics must also realize the required conditions during execution. Notably, reported fidelity re- mains unchanged from B5 to AURORA for all five models, while verified-pass rates increase. The repair gains therefore do not coincide with a broad reduction in specification fidelity, although individual edits still require inspection when thresholds or explicit conditions are modified. Fidelity and verification thus provide complementary evidence of requirement preservation and realized behavior. Because B1– B3 lack executable specifications for semantic verification, their completion rates alone cannot establish whether the generated scenarios fulfill the requests. C. Grounding and Execution Validation We next examine scenario grounding and coordinated execution across the coupled simulators. a) Spatial grounding and mission realization: Fig. 10 shows close agreement between requested and realized sce- nario geometry. Ground-agent placements are generally accu- rate, with larger deviations mainly caused by projection onto valid sidewalks or lanes. UAV missions likewise remain close to their intended geometry, with navigation deviations con- centrated around turns. Most geometric differences therefore reflect feasibility enforcement rather than grounding failure. Tracking adds a temporal dimension: valid initialization does 050 Requested Distance (m) 0 20 40 60 80 Achieved (m) a Agent Placement 0510 Placement Error (m) pedestrian police bus van motorcycle truck b Error by Agent Type Score (per-term min–max) comm. sensing airspace road joint 0.70.91 0.40.75 0.891 0.251 0.860.95 verifiedsilent c Scores 02040 Time (s) 0.0 0.5 1.0 Altitude / Requested requested altitude d Hover: Altitude Hold 01020 Distance from Planned Route (m) 0% 10% 20% 30% 40% Share of Samples median 1.7 m e Navigate: Route 02040 Time Since Tracking Start (s) 0 20 40 60 Distance to Target (m) tracking threshold f Track: Convergence Fig. 10. Grounding validation: placement accuracy and per-type error (a, b), anchor scores by outcome (c), and mission realization (d–f). not ensure immediate mission fulfillment, because the UAV must first converge to the moving target and maintain the required relation over time. Placement accuracy and mis- sion realization consequently answer different questions— whether the scenario is grounded correctly at initialization and whether the requested relation persists once the agents begin to move. b) Grounding quality and scenario outcomes: Verified runs have higher joint anchor scores than silent failures, with the largest separation in the sensing term (Fig. 10). Road, airspace, and communication scores differ less between the two groups. Successful grounding therefore depends not only on valid roads and aerial clearance, but also on whether the selected location supports the sensing relationship required by the interaction. Joint grounding captures these constraints before execution, while the resulting traces reveal whether the anticipated relationships persist as agents move. The an- chor score is therefore useful as a pre-execution indicator of scenario quality, but it cannot replace trace-level verification of the realized interaction. c) Coordination across models: Fig. 11 shows con- sistent simulator coordination across the five LLMs, with small clock offsets and limited event-timing error. The distributions nevertheless include delayed events and larger placement deviations, making variation beyond the median important. Mission performance varies more substantially, particularly in tracking convergence and time on target. Because all models use the same managed executor, this contrast separates low-level co-simulation reliability from model-dependent scenario realization. Stable synchronization provides the foundation for coordinated execution, but does not by itself guarantee the requested air–ground behavior. D. Repair Analysis and Computational Cost We further analyze repair effectiveness and its computa- tional cost. a) Repair effectiveness: Repair connects violated con- ditions to the components that can be adjusted. In a tracking example, two iterations move the tracked-fraction margin from negative to positive, while infeasible patrol routes are 9 minmax PHYSICAL ↓ Agent placement error (m) 013.6 ↓ Simulator clock offset (ms) 3.58.7 TEMPORAL ↓ Event timing error (s) 056.3 MISSION & SAFETY ↓ Drone catch-up time (s) 040 ↑ Time on target (fraction) 01 ↓ Tracking distance error (m) 2.730.8 ↑ Min. time-to-collision (s) 0.3965 ↑ Closest agent distance (m) 3.378 Gemini 3.1 ProGPT-4oGemini 3.8 FlashGPT-5.4 MiniGPT-5.5 Fig. 11. Execution and mission metrics across language models. LLM parsing Scene setup Scenario execution 10 0 10 1 10 2 Wall-Clock Time (s, Log Scale) a Pipeline Stages Repair episode Simulator re-run / iter. Diagnosis & edits / iter. b Closed-Loop Repair Gemini 3.1 ProGPT-4oGemini 3.8 FlashGPT-5.4 MiniGPT-5.5 Fig. 12. Computational cost across models: (a) pipeline-stage latency; (b) repair-episode cost and per-iteration components. corrected during grounding through waypoint snapping. The two cases illustrate a useful division of responsibility: local- ized quantitative violations are handled by the repair loop, whereas structural geometric defects are returned to ground- ing rather than forced through small parameter edits. Routing failures to the appropriate stage avoids regenerating the full scenario while keeping each correction within its intended scope. Improved margins must nevertheless be interpreted alongside the edits that produced them. Some repairs relax acceptance thresholds, and parsing may reformulate an in- feasible request instead of rejecting it. Provenance weighting discourages such changes but does not prohibit them. Repair logs therefore distinguish corrections to realized behavior from changes to the conditions used to evaluate it. Verified- pass rate should consequently be interpreted together with prompt fidelity and edit history rather than as a standalone measure of preservation of the original request. b) Computational cost: Fig. 12 shows that simulator execution dominates repair cost. Median repair episodes take 194–249 s across models, largely due to repeated co- simulation runs, while diagnosis and edit generation con- tribute comparatively little. Parsing latency varies across LLMs, whereas static checks and kinematic preview remain inexpensive. This asymmetry supports placing inexpensive verification and surrogate screening before full re-execution whenever possible. Localized edits constrain the search space, but the larger opportunity for reducing repair cost lies in avoiding unnecessary simulator reruns rather than accelerating edit generation alone. VI. CONCLUSION AURORA connects natural-language scenario generation with explicit feasibility checking, synchronized air–ground execution, and trace-based verification. Its Air–Ground Sce- nario Graph links requested relations to executable condi- tions, execution evidence, and repair variables, providing a shared representation for grounding, monitoring, failure localization, and refinement. Evaluation across five language models distinguishes three properties of generated scenarios: whether they execute, preserve the request, and realize the requested conditions. Structured execution improves relia- bility, but successful completion can still mask semantic failures; runtime verification exposes these failures, while localized repair resolves many without regenerating the full scenario. More broadly, the results support treating language- driven scenario generation as verified compilation rather than direct code generation. Explicit intermediate representations make cross-domain dependencies inspectable and provide a structured path from runtime violations to the parameters that can affect them, reducing reliance on language models for simulator-specific reasoning and improving failure diagnosis. The current framework supports a defined scenario schema, three UAV mission templates, simplified sensing and communication models, and one simulation environment. Because parsing can reinterpret infeasible requests and repair can relax monitored thresholds, a verified pass should be interpreted as satisfaction of the encoded monitored condi- tions rather than unconditional satisfaction of the original prompt. Future work will strengthen preservation of explicit requirements and terminal handling of infeasible requests, while extending AURORA toward multi-UAV planning and control, richer cooperative air–ground interactions, vision– language–action models for closed-loop perception and de- cision making, and additional simulation environments and real-world testbeds. REFERENCES [1] T. Feng, X. Wang, F. Han, L. Zhang, and W. Zhu, “U2udata: A large-scale cooperative perception dataset for swarm uavs autonomous flight,” in Proceedings of the 32nd ACM International Conference on Multimedia, 2024, p. 7600–7608. [2] F. Yang, B. Yu, Y. Zhou, X. Luo, Z. Tu, and C. Liu, “Edge-based multimodal sensor data fusion with vision-language-action (vla) model for real-time autonomous vehicle accident avoidance,” Engineering Applications of Artificial Intelligence, vol. 180, p. 115186, 2026. [3] R. Xu, H. Xiang, Z. Tu, X. Xia, M.-H. Yang, and J. Ma, “V2x-vit: Vehicle-to-everything cooperative perception with vision transformer,” in European conference on computer vision. Springer, 2022, p. 107– 124. [4] K. Wu, P. Li, Y. Cheng, S. T. Parker, B. Ran, D. A. Noyce, and X. Ye, “A digital twin framework for physical-virtual integration in v2x- enabled connected vehicle corridors,” IEEE Transactions on Intelligent Transportation Systems, vol. 26, no. 9, p. 14 407–14 420, 2025. [5] X. Gao, Y. Wu, R. Wang, C. Liu, Y. Zhou, and Z. Tu, “Langcoop: Collaborative driving with language,” in 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, 2025, p. 4226–4237. 10 [6] X. Gao, Y. Wu, F. Yang, X. Luo, K. Wu, X. Chen, Y. Wang, C. Liu, Y. Zhou, and Z. Tu, “Airv2x: Unified air-ground vehicle-to-everything collaboration,” arXiv preprint arXiv:2506.19283, 2025. [7] T. Zeng, Y. Wen, and H. Zhang, “Carla-air: Fly drones inside a carla world–a unified infrastructure for air-ground embodied intelligence,” arXiv preprint arXiv:2603.28032, 2026. [8] H. Zhang, X. Yue, K. Tian, S. Li, K. Wu, Z. Li, D. Lord, and Y. Zhou, “Virtual roads, smarter safety: A digital twin framework for mixed autonomous traffic safety analysis,” IEEE Internet of Things Journal, 2025. [9] Y. Wang, S. Xing, C. Cui, R. Li, H. Hua, K. Tian, Z. Mo, X. Gao, M. Pavone, Y. Zhou, et al., “Generative ai for autonomous driving: Frontiers and opportunities,” ACM Computing Surveys, 2025. [10] K. Wu, P. Li, Y. Zhou, R. Gan, J. You, Y. Cheng, J. Zhu, S. T. Parker, B. Ran, D. A. Noyce, et al., “V2x-llm: Enhancing v2x integration and understanding in connected vehicle corridors,” arXiv preprint arXiv:2503.02239, 2025. [11] X. Gao, K. Wu, H. Zhang, K. Tian, Y. Zhou, and Z. Tu, “Automated vehicles should be connected with natural language,” arXiv preprint arXiv:2507.01059, 2025. [12] B.-K. Ruan, H.-T. Tsui, Y.-H. Li, and H.-H. Shuai, “Traffic scene generation from natural language description for autonomous vehicles with large language model,” arXiv preprint arXiv:2409.09575, 2024. [13] J. Zhang, C. Xu, and B. Li, “Chatscene: Knowledge-enabled safety- critical scenario generation for autonomous vehicles,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2024, p. 15 459–15 469. [14] C. Chang, S. Wang, J. Zhang, J. Ge, and L. Li, “Llmscenario: Large language model driven scenario generation,” IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 54, no. 11, p. 6581– 6594, 2024. [15] Z. Sheng, Z. Huang, Y. Qu, Y. Leng, and S. Chen, “Talk2traffic: Interactive and editable traffic scenario generation for autonomous driving with multimodal large language model,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, p. 3827–3836. [16] A. Phadke, A. Hadimlioglu, T. Chu, and C. N. Sekharan, “Integrating large language models for uav control in simulated environments: A modular interaction approach,” arXiv preprint arXiv:2410.17602, 2024. [17] A. Shibu, M. Saleh, M. Al-Musleh, and N. Abdulaziz, “Skysim: A ros2-based simulation environment for natural language control of drone swarms using large language models,” arXiv preprint arXiv:2602.01226, 2026. [18] A. Koubaa and K. Gabr, “Agentic uavs: Llm-driven autonomy with integrated tool-calling and cognitive reasoning,” arXiv preprint arXiv:2509.13352, 2025. [19] T. Dreossi, D. J. Fremont, S. Ghosh, E. Kim, H. Ravanbakhsh, M. Vazquez-Chanlatte, and S. A. Seshia, “Verifai: A toolkit for the formal design and analysis of artificial intelligence-based systems,” in International Conference on Computer Aided Verification. Springer, 2019, p. 432–442. [20] C. Xu, W. Ding, W. Lyu, Z. Liu, S. Wang, Y. He, H. Hu, D. Zhao, and B. Li, “Safebench: A benchmarking platform for safety evaluation of autonomous vehicles,” Advances in Neural Information Processing Systems, vol. 35, p. 25 667–25 682, 2022. [21] J. Wang, X. Cao, J. Zhong, Y. Zhang, Z. Han, H. Yu, C. Zhang, L. He, S. Xu, and J. Wang, “Griffin: Aerial-ground cooperative detection and tracking dataset and benchmark,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 12, 2026, p. 9867– 9875. [22] H. Ye, R. Sunderraman, and S. Ji, “Uav3d: A large-scale 3d percep- tion benchmark for unmanned aerial vehicles,” Advances in Neural Information Processing Systems, vol. 37, p. 55 425–55 442, 2024. [23] Y. Cui, X. Dong, B. Gao, J. Xiang, D. Li, and Z. Tu, “Airsimag: A high-fidelity simulation platform for air-ground collaborative robotics,” arXiv preprint arXiv:2603.23079, 2026. [24] R. Xu, H. Xiang, X. Xia, X. Han, J. Li, and J. Ma, “Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to- vehicle communication,” arXiv preprint arXiv:2109.07644, 2021. [25] H. Yu, Y. Luo, M. Shu, Y. Huo, Z. Yang, Y. Shi, Z. Guo, H. Li, X. Hu, J. Yuan, et al., “Dair-v2x: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, p. 21 361–21 370. [26] J. You, Z. Jiang, Z. Huang, H. Shi, R. Gan, K. Wu, X. Cheng, X. Li, and B. Ran, “V2x-vlm: End-to-end v2x cooperative autonomous driv- ing through large vision-language models,” Transportation Research Part C: Emerging Technologies, vol. 183, p. 105457, 2026. [27] A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V. Koltun, “Carla: An open urban driving simulator,” in Conference on robot learning. PMLR, 2017, p. 1–16. [28] S. Shah, D. Dey, C. Lovett, and A. Kapoor, “Airsim: High-fidelity visual and physical simulation for autonomous vehicles,” in Field and service robotics: Results of the 11th international conference. Springer, 2017, p. 621–635. [29] R. Xu, Y. Guo, X. Han, X. Xia, H. Xiang, and J. Ma, “Opencda: an open cooperative driving automation framework integrated with co-simulation,” in 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021, p. 1155–1162. [30] D. J. Fremont, T. Dreossi, S. Ghosh, X. Yue, A. L. Sangiovanni- Vincentelli, and S. A. Seshia, “Scenic: a language for scenario specification and scene generation,” in Proceedings of the 40th ACM SIGPLAN conference on programming language design and imple- mentation, 2019, p. 63–78. [31] ASAM e.V., “ASAM OpenSCENARIO: Dynamic content in driv- ing simulation, standard v1.2,” https://w.asam.net/standards/detail/ openscenario/, 2022. [32] Z. Zhong, D. Rempe, D. Xu, Y. Chen, S. Veer, T. Che, B. Ray, and M. Pavone, “Guided conditional diffusion for controllable traffic simulation,” arXiv preprint arXiv:2210.17366, 2022. [33] K. Wu, Y. Zhou, H. Shi, D. Lord, B. Ran, and X. Ye, “Hypergraph- based motion generation with multi-modal interaction relational rea- soning,” Transportation Research Part C: Emerging Technologies, vol. 180, p. 105349, 2025. [34] R. Gan, P. Li, K. Long, B. An, J. You, K. Wu, and B. Ran, “Plan- ning safety trajectories with dual-phase, physics-informed, and trans- portation knowledge-driven large language models,” arXiv preprint arXiv:2504.04562, 2025. [35] S. Tan, B. Ivanovic, X. Weng, M. Pavone, and P. Kraehen- buehl, “Language conditioned traffic generation,” arXiv preprint arXiv:2307.07947, 2023. [36] Q. Li, Z. M. Peng, L. Feng, Z. Liu, C. Duan, W. Mo, and B. Zhou, “Scenarionet: Open-source platform for large-scale trafficscenariosimulationandmodeling,”inAdvancesin Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol.36.CurranAssociates,Inc.,2023,p.3894–3920. [Online]. Available: https://proceedings.neurips.c/paper files/paper/ 2023/file/0c26a501df8fb919a0350e2df06b5d39-Paper-Datasetsand Benchmarks.pdf [37] M. Chen, J. Tworek, H. Jun, Q. Yuan, H. P. D. O. Pinto, J. Ka- plan, H. Edwards, Y. Burda, N. Joseph, G. Brockman, et al., “Evaluating large language models trained on code,” arXiv preprint arXiv:2107.03374, 2021. [38] J. Liang, W. Huang, F. Xia, P. Xu, K. Hausman, B. Ichter, P. Florence, and A. Zeng, “Code as policies: Language model programs for embodied control,” in 2023 IEEE International conference on robotics and automation (ICRA). IEEE, 2023, p. 9493–9500. [39] S. Yao, J. Zhao, D. Yu, I. Shafran, K. R. Narasimhan, and Y. Cao, “React: Synergizing reasoning and acting in language models,” in NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. [40] W. Wang, Y. Li, L. Jiao, and J. Yuan, “Large language model-driven closed-loop uav operation with semantic observations,” IEEE Internet of Things Journal, 2025. [41] R. Dechter, I. Meiri, and J. Pearl, “Temporal constraint networks,” Artificial intelligence, vol. 49, no. 1-3, p. 61–95, 1991. [42] O. Maler and D. Nickovic, “Monitoring temporal properties of con- tinuous signals,” in International symposium on formal techniques in real-time and fault-tolerant systems. Springer, 2004, p. 152–166. [43] G. E. Fainekos and G. J. Pappas, “Robustness of temporal logic spec- ifications for continuous-time signals,” Theoretical Computer Science, vol. 410, no. 42, p. 4262–4291, 2009. [44] A. Donz ́ e and O. Maler, “Robust satisfaction of temporal logic over real-valued signals,” in International conference on formal modeling and analysis of timed systems. Springer, 2010, p. 92–106. 11

Instant research discovery